中国农业科学 ›› 2019, Vol. 52 ›› Issue (1): 129-142.doi: 10.3864/j.issn.0578-1752.2019.01.012

• 食品科学与工程 • 上一篇    下一篇

基于BP人工神经网络算法的苹果制干适宜性评价

张彪(),刘璇(),毕金峰,吴昕烨,金鑫,李旋,李潇   

  1. 中国农业科学院农产品加工研究所/农业农村部农产品加工重点实验室,北京 100193
  • 收稿日期:2018-05-29 接受日期:2018-09-18 出版日期:2019-01-01 发布日期:2019-01-12
  • 通讯作者: 刘璇
  • 基金资助:
    “十三五”国家重点研发计划(2016YFD0400201-4)

Suitability Evaluation of Apple for Chips-Processing Based on BP Artificial Neural Network

ZHANG Biao(),LIU Xuan(),BI JinFeng,WU XinYe,JIN Xin,LI Xuan,LI Xiao   

  1. Institute of Food Science and Technology, Chinese Academy of Agricultural Sciences/Key Laboratory of Agro-Products Processing, Ministry of Agriculture and Rural Affairs, Beijing 100193
  • Received:2018-05-29 Accepted:2018-09-18 Online:2019-01-01 Published:2019-01-12
  • Contact: Xuan LIU

摘要:

【目的】 建立苹果原料制干适宜性评价模型,实现基于苹果原料指标预测干制品品质的目标,为苹果制干专用化原料的筛选提供方法依据,为明确苹果干制品品质形成的基础物质提供数据支持。【方法】 以来自7个不同主产区的21个主栽品种,共34份苹果鲜果样本为研究对象,运用多种数据处理方法建立苹果脆片品质综合评价模型与苹果原料制干适宜性评价模型。(1)利用压差闪蒸干燥方法制备34份苹果鲜果的脆片样本,测定苹果脆片17项品质指标,采用因子分析进行降维并筛选得到苹果脆片品质评价核心指标,运用层次分析法得到脆片核心指标权重值,构建脆片品质综合评价模型并计算得到脆片综合评价得分。(2)测定34份苹果鲜果样本22项品质指标,与脆片核心指标进行相关性分析并筛选出与脆片品质相关的果实特征指标。选用29个样本以果实特征指标为输入,对应脆片综合评价得分为输出,利用误差反向传播(Error Back Propagation, BP)神经网络算法构建学习模型;其余5个样本为验证样本,评价学习模型的预测准确性。变换3组学习样本构建3个学习模型,对比3个模型的预测准确性,验证建模方法的合理性与稳定性。【结果】 苹果脆片L*值、脆度、膨化度、可滴定酸含量、可溶性糖含量和粗蛋白含量被确定为不同样本脆片品质综合评价的核心指标,构建的苹果脆片品质综合评价模型为Y综合得分=L*值×0.3724+脆度×0.2665+膨化度×0.1583+可滴定酸含量×0.0890+可溶性糖含量×0.0569+粗蛋白含量×0.0569。34个苹果鲜果样本制得的脆片综合得分范围为0.2069—0.7933,存在较大差异,得分排名前3的苹果样本为‘辽宁华红’‘辽宁华金’和‘山东烟富6号’,排名最后的苹果样本为‘陕西秦冠’。基于脆片核心指标与苹果果实品质指标相关性分析结果,筛选出苹果果实的果形指数、果肉a*值、pH、可滴定酸含量、Vc含量、果核比例、粗蛋白含量、果肉b*值、密度、可溶性固形物含量、粗纤维含量、总糖含量12项指标作为果实制干适宜性评价的特征指标。以果实特征指标值为输入层,对应苹果脆片综合评分为输出层,建立BP神经网络学习模型,可实现苹果原料制干适宜性的定量预测。该方法建立的学习模型有较高的预测准确性与稳定性,变换学习样本得到的3个学习模型的预测值与实际值相对误差均不超过10%,实际值与模型预测值线性拟合后决定系数R 2均大于0.95。【结论】 苹果制干适宜性可由果实的果形指数、果肉a*值、pH、可滴定酸含量、Vc含量、果核比例、粗蛋白含量、果肉b*值、密度、可溶性固形物含量、粗纤维含量、总糖含量12项指标进行评价,建立的适宜性评价模型可实现基于苹果原料指标定量预测其制干适宜性。

关键词: 苹果, 脆片, 干制, 适宜性评价, BP神经网络

Abstract:

【Objective】The aim of the paper was to establish suitability evaluation model for apple chips-processing from different cultivars and to achieve the quality prediction of apple chips based on raw material indicators.【Method】34 fresh apple samples of 21 apple varieties from 7 major growing regions were selected as research objects. Factor analysis (FA) and analytic hierarchy process (AHP) were used to establish comprehensive quality evaluation model for chips, and Error Back Propagation (BP) artificial neural network was used to establish chips-processing suitability evaluation model for apple fruits. (1) Chips were prepared by instant controlled pressure drop (DIC, French for détente instantannée controlee, also known as explosion puffing) and 17 indicators were measured. The core indexes of chips were selected by FA and correlation analysis. The weights of the core indexes were determined by AHP, and then the comprehensive quality evaluation scores of chips were calculated. (2) 22 indicators of 34 fruit samples with different cultivars and regions were measured. Then the characteristic indicators of apple fruits related to chip qualities were screened out by correlation analysis between data groups of apple fruit indicators and chip core indexes. Learning model with input of fruit characteristic indicators and output of chip comprehensive evaluation scores was established by database of 29 apple samples. 5 apple samples were chosen as test samples to verify the prediction accuracy of the learning model. Modified leaning models from different sample groups were compared by prediction accuracy, which could be the evidence to evaluate rationality and stability for application of BP neural network in the present research.【Result】The results showed that L* value, brittleness, puffing degree, titratable acid, soluble sugar and crude protein of apple chip were determined as the core indexes which the weights were 0.3724, 0.2665, 0.1583, 0.0890, 0.0569 and 0.0569, respectively. The comprehensive quality scores of chips from 34 apple samples ranged from 0.2069 to 0.7933, indicating significant variation. The top 3 apple samples with high scores were Liaoning Huahong, Liaoning Huajin and Shandong Yanfu 6, and the final ranking for Shanxi Qinguan. Correlation analysis was performed between core indexes of chips and quality indicators of apple raw materials to achieve characteristic indicators of apple fruits, including the fruit shape index, a* value (pulp), pH value, titratable acid content, Vc content, proportion of core, protein content, b* value (pulp), density, soluble solids content, crude fiber content and total sugar content. Therefore, learning models were established with input layer of the characteristic indicators value of fruit and output layer of the comprehensive quality score of apple chip, which could predict the comprehensive quality of apple chips from indicators of raw materials. Moreover, the model showed high prediction accuracy. The relative errors between the predicted and actual values of the three learning models groups did not exceed 10%, and the coefficients of determination R 2 of linear fitting were higher than 0.95.【Conclusion】Suitability evaluation of apple fruit for chips-processing could be evaluated by fruit shape index, a* value (pulp), pH value, titratable acid content, Vc content, proportion of core, protein content, b* value (pulp), density, soluble solids content, crude fiber content and total sugar content. The established model could be used to quantitatively predict apple fruit suitability for chips-processing based on the indicators of raw fruits.

Key words: apple, chips, dehydration, suitability evaluation, BP neural network